Source-linked AI summary
LoaDiff: Conditional Generation of Electricity Consumption Time Series for Energy Analytics
Mariia Baranova, Adrien Petralia, Etienne Le Naour, Nathan Etourneau, Guillaume Hofmann, Themis Palpanas
TL;DR
Restricted access to granular smart-meter data creates a need for realistic synthetic alternatives that preserve useful household electricity-consumption information. LOADIFF uses conditional diffusion to generate year-long, sub-hourly load curves from static household and dynamic contextual variables, and it performs strongly across fidelity, privacy, downstream utility, and controllability evaluations. The results support its use as a practical privacy-preserving substitute, while the privacy and temperature analyses remain limited in scope.
Problem
Smart-meter load curves are personal data, motivating synthetic generation that preserves realistic electricity-consumption structure without corresponding to real households.
Method
LOADIFF is a conditional diffusion model that generates year-long, sub-hourly load curves using static household attributes and dynamic contextual variables.
Results
Across three datasets, LOADIFF ranks among the top methods for fidelity and privacy, provides the strongest downstream utility, and wins 16 of 18 classifier-appliance combinations.
Takeaways & Limitations
LOADIFF provides a practical privacy-preserving substitute for individual smart-meter data while preserving information useful for forecasting and appliance detection.
Takeaways & Limitations
Privacy assessment uses distance-based heuristics rather than formal guarantees or attack-based evaluations, and temperature sensitivity covers only two counterfactual shifts.
Abstract
from arXiv · showhide
The energy transition is reshaping residential electricity consumption through the increasing adoption of distributed generation, electrified appliances, and demand-response programs. Understanding these evolving behaviors requires access to granular smart-meter data for applications such as load forecasting, appliance detection, and demand-side flexibility analysis. However, such data are subject to strict access restrictions and data-protection regulations. Thus, realistic synthetic alternatives are necessary. In this paper, we introduce LoaDiff, a diffusion-based generative model for year-long, sub-hourly smart-meter load curves. LoaDiff supports flexible conditioning on static household attributes, such as appliance ownership, and dynamic contextual variables, including calendar information and outdoor temperature. We evaluate the model against multiple generative baselines on three residential electricity-consumption datasets. Our experiments assess four complementary dimensions: fidelity and diversity, training-record memorization risk, downstream utility for load forecasting and appliance detection, and conditional controllability under alternative temperature conditions. The results show that LoaDiff generates realistic and diverse load profiles, achieves a favorable trade-off between generation quality and limited evidence of memorization, preserves information useful for downstream energy applications, and responds coherently to changes in conditioning variables.
I. INTRODUCTION
LOADIFF addresses restricted access to granular household electricity data by generating year-long, sub-hourly load curves with conditioning on household and contextual variables. It is evaluated across fidelity, privacy risk, downstream utility, and controllability.
- Smart-meter load curves are personal data under GDPR, restricting access because they can reveal occupancy schedules, work habits, holidays, or appliance use.
- Existing generative approaches often target short horizons, offer limited control over weather, calendars, or appliance ownership, and lack systematic evaluation across fidelity, utility, and privacy risk.
- LOADIFF generates year-long, sub-hourly load traces conditioned on static household attributes and dynamic variables such as calendar information and outdoor temperature.
- Across three datasets, the evaluation examines fidelity and diversity, memorization risk, downstream forecasting and appliance-detection utility, and conditional controllability.
II. RELATED WORK
Related work spans time-series generative modeling, conditional controllability, and electricity-load synthesis. The paper frames its evaluation around fidelity, practical utility, and empirical privacy risk.
- Classical ARIMA, Gaussian Mixture, and hidden Markov models capture seasonal patterns but struggle with complex nonlinear dynamics and heterogeneous behaviors.
- Conditional generation incorporates static covariates or dynamic exogenous variables because unconditional synthesis often lacks control over specific attributes.
- LOADIFF formalizes load curves as long-horizon sub-hourly series conditioned on static descriptors s and dynamic covariates z.
- Synthetic smart-meter data are evaluated by fidelity, utility, and empirical privacy risk rather than fidelity metrics alone.
IV. THE LOADIFF APPROACH
LOADIFF uses conditional diffusion with a DiT-like Transformer to generate year-long load curves efficiently while preserving daily structure and integrating covariates.
- LOADIFF progressively removes noise from corrupted time series during training and iteratively denoises Gaussian noise at inference to generate conditioned load curves.
- Year-long half-hourly curves are reshaped into daily patches, producing 365 tokens so the Transformer can model long horizons without prohibitively long raw sequences.
- Each patch combines consumption values, aligned dynamic covariates, and interaction features before projection into the Transformer embedding space.
- Calendar-based positional encoding represents weekday, day-of-month, day-of-year, and month information using periodic representations.
- The denoiser uses stacked Transformer blocks with multi-head self-attention, feed-forward networks, residual connections, and AdaLN conditioning.
C. Conditioning on exogenous factors
LOADIFF conditions generation on exogenous factors to support scenario-based and counterfactual analyses. Its architecture combines long-range temporal modeling with controllable synthetic load generation.
- Static household descriptors such as heater or electric-vehicle ownership are embedded and used to condition generation on specified household scenarios.
- The conditioning design supports scenario and counterfactual generation involving factors such as heating technology and weather year while avoiding overfitting to spurious correlations.
- The DiT-like architecture jointly models long-range temporal dependencies, integrates exogenous information, and generates controllable year-long synthetic load curves.
A. Datasets
The study evaluates LOADIFF on three residential electricity-consumption datasets, including real household data and a simulator-backed synthetic dataset, under multiple conditioning regimes and generative baselines.
- Three residential electricity-consumption datasets are used: one from ISSDA and two provided by EDF.
- The CER residential subset contains 4,225 households recorded every 30 minutes, with household-composition and appliance-ownership questionnaires.
- EDF 1 contains 2,083 houses with 30-minute consumption recordings and household appliance information, while EDF 2 provides 20K simulator-generated customers conditioned on household, appliance, and weather variables.
- The evaluation compares LOADIFF with generative models across unconditional, static-conditioned, and hybrid-conditioned regimes supported by each method.
2) Evaluation Metrics:
The evaluation measures generated-data fidelity, privacy risk, downstream utility, and conditional controllability using complementary metrics, task protocols, and standardized experimental settings.
- The experiments assess fidelity and diversity, privacy, downstream utility, and controllability under alternative conditioning scenarios.
- Fidelity and diversity: Fidelity uses Discriminative1NN, ROCKET-based Fréchet distance, and autocorrelation discrepancy to compare synthetic and real load distributions.
- Privacy: Privacy is assessed with NNDR and NeighborsPrivacy, whose 0.5 reference indicates balanced real/synthetic neighborhoods and limited memorization evidence.
- Downstream utility: Downstream utility compares TRTR, TSTR, and TR+STR protocols for load forecasting and appliance detection.
- Conditional controllability: Controllability is tested through counterfactual temperature-sensitivity analysis, while models use identical 70/15/15% train/validation/test splits and normalized data.
C. Results
Across datasets, LOADIFF and EnergyDiff provide the strongest distributional alignment, while privacy results show no consistent quality–memorization trade-off and t-SNE overlap guides visual comparison.
- Data Generation Quality: Figure 3 compares real and generated distributions across baselines and datasets, where greater t-SNE overlap indicates better synthetic coverage of the real-data manifold.
- Data Generation Quality: GMM only partially recovers real-distribution modes despite a perfect Discriminative1NN score, consistent with weaker FID and ACD performance.
- Data Generation Quality: LOADIFF and EnergyDiff jointly achieve the best and second-best FID and ACD scores on every dataset, outperforming VAE- and GAN-based methods on distributional and temporal alignment.
- Data Generation Quality: LOADIFF ranks best or second-best on both FID and ACD across every dataset, although its Discriminative1NN score varies from 0.5104 on CER to mid-table on EDF 1 and EDF 2.
- Privacy: LOADIFF has the lowest NNDR on CER (0.0009) and EDF 2 (0.0005), while EnergyDiff is lowest on EDF 1 (0.0013), showing no consistent quality–memorization trade-off.
2) Downstream Utility:
LOADIFF preserves useful temporal and appliance-specific information for downstream forecasting and detection, often outperforming competing generators and benefiting from combination with real data.
- Forecasting: Across most datasets and protocols, LOADIFF outperforms competing baselines in forecasting, with its best overall performance in the REAL + SYNTH. setting.EnergyDiff and LOADIFF achieve forecasting performance comparable to or better than the real-data-only reference.
- Appliance detection: LOADIFF ranks first overall for appliance detection, achieving average balanced accuracy of 0.710/0.766 and average rank of 1.67/2.44 with ROCKET / TransApp.The strongest competing average scores are 0.660/0.677 for EnergyDiff and 0.631/0.683 for GMM.
- Appliance detection: LOADIFF is best among generators in 16 of 18 classifier-appliance combinations under synthetic-only training.On EDF2, synthetic-only balanced accuracies reach 0.907/0.942 for electric heating, 0.621/0.874 for air conditioning, and 0.846/0.864 for water heating.
- Appliance detection: Adding LOADIFF samples to real training improves performance in 12 of 18 classifier-appliance combinations, including all six EDF2 cases.The gains are not systematic across every dataset and classifier, but support augmentation beyond synthetic-only replacement.
- Appliance detection: For electric heating on EDF2 and cookers on CER, LOADIFF most closely reproduces average daily profiles and appliance-specific class differences.This qualitative comparison suggests that conditioning preserves meaningful appliance signatures.
3) Conditional Controllability: Temperature Sensitivity:
LOADIFF responds coherently to counterfactual winter-temperature changes in generated yearly profiles. Colder conditions increase electric-heating demand, while warmer conditions reduce it asymmetrically.
- Temperature sensitivity: A −5 °C winter shift raises mean winter load from 3,395.4 W to 3,754.4 W (+10.6%), while a +5 °C shift lowers it to 1,727.7 W (−49.1%).The experiment enables the electric-heating label for all generated samples and compares the reference profile with two counterfactual temperature scenarios.
- Temperature sensitivity: The asymmetric load response is consistent with a nonlinear relationship between temperature and electric-heating demand.Characterizing the precise response shape would require evaluating additional intermediate temperature conditions.
D. Ablations
The reference LOADIFF configuration offers the strongest overall ablation trade-off, while conditioning choices and temporal patch size materially affect generation quality. The full-method conclusion remains strong across datasets, but performance varies by dataset and metric.
- The reference configuration achieves the best overall trade-off, with FID 0.0128, ACD 0.0086, and Disc1NN 0.5175.It has the lowest FID and ACD and a near-optimal Disc1NN score among the ablations.
- Removing temperature conditioning most severely degrades the metric profile, while removing calendar features causes a milder but consistent decline.Removing static appliance labels leaves FID essentially unchanged but worsens Disc1NN and ACD.
- Replacing AdaLN conditioning or disabling classifier-free guidance degrades all three metrics relative to the reference.Without classifier-free guidance, ACD worsens most substantially and Disc1NN also declines.
- Temporal patch size has the largest ablation impact: weekly patches raise FID 2.7× and ACD 6.5× despite achieving Disc1NN 0.5000.The reference uses one patch per calendar day, matching the load curves’ diurnal periodicity.
- Across three datasets, LOADIFF ranks among the top methods for fidelity and privacy and wins 16 of 18 classifier–appliance combinations, though no method dominates every metric.The authors therefore emphasize evaluating multiple complementary metrics rather than relying on a single score.
- The study is constrained by proprietary datasets, heuristic privacy assessment, and temperature sensitivity tested under only two counterfactual shifts.These limitations make the privacy and temperature findings indicative rather than exhaustive.